Auto permission mode doesn't make Claude Code safe
Auto mode lowers the entry bar, but users stop learning the permission system and how to isolate their environment. Why manual mode, oddly enough, pushed people towards safety.
// writing
Posts
Notes on AI adoption, agentic engineering, architecture, hiring and engineering leadership. Originally published in my Telegram channel and on LinkedIn.
In defence of the token-spend metric: AI adoption vs delivery
Token spend is a poor measure of value, but a good measure of adoption. Why an AI rollout needs two stages, each with its own metrics.
Nobody to configure JIRA: how the Scrum Master role got devalued
From the Agile Manifesto and digital transformations to 'we have nobody to configure JIRA'. How the Scrum Master role lost its meaning.
Do we still need to watch conference talks in the age of AI?
Inspiration, broader horizons, new knowledge: which reasons to watch conference talks still hold when AI assistants can summarise, research and file knowledge for you?
Agent trends: knowledge management and harness engineering
Long-term knowledge management and secure, full-featured agent environments are the next step towards autonomous AI agents. Lessons from OpenAI, Karpathy and Anthropic.
Reality check: Claude Code's own bug tracker
Demos show Claude Code triaging bugs, fixing them test-first and guarding against regressions. Its own GitHub repository tells a different story.
Scaling AI across an organisation: seven things that matter
Memes about enterprise AI adoption are funny because they ring true. A saner approach, in order of importance: from a strategy everyone can explain to honest measurement.
Why I believe in autonomous agents: a ChatGPT agent case
Asked to chart IronMan 70.3 finish times, ChatGPT agent clicked through a chart, read the page code, wrote Python and cross-checked its own data. Persistence many people would envy.
Why AI slowed down experienced developers in the METR study
The headline says AI tools made developers 19% slower. Read how the experiment was set up and the result is expected: without formalised project knowledge, experts are faster by hand.
Layoffs at record profits: three ways to lower your risk
Even profitable companies no longer keep people 'just in case'. Continuous learning, visible use of AI and taking part in transformation raise your chances of staying.
2025, the year of autonomous coding agents
From Devin to Codex and Jules: why software development became the first big arena for autonomous AI agents, and what the next step in their evolution looks like.
People, not tools, are what holds back AI adoption
The tools get better every month. What's missing is an AI-driven mindset, and organisations that don't see results are going back to measuring individual performance.
Using AI to analyse job interviews
A full interview transcript and a model with a large context window give you candidate feedback, interviewer coaching and a more objective assessment. A game-changing use of AI.
Four kinds of seniority, and which one you are hiring for
Technological, corporate, applied and domain seniority are all real, and each is valuable in its own context. Effective hiring starts with knowing which mix you need.
Three traps when you step up to the next level of influence
Moving from developer to tech lead, or from tech lead to architect? Operational work, the pull of easy tasks and distrust of other people's work are the traps to get past fast.
CQRS and event-driven architecture are not the same thing
People often treat CQRS and event-driven architecture as one idea, but the link between them is weak. What each is really about, and why neither needs microservices.
// next step
Let’s talk about your AI delivery gap
A 30-minute intro call: you tell me where you are, I tell you honestly whether and how I can help.